一种基于强化学习的自适应多邻域人工蜂群算法  

Adaptive Multi-Neighborhood Artificial Bee Colony Algorithm Based on Reinforcement Learning

作　　者：周新宇[1] 尹子悦 高卫峰 谭贵森 易玉根 ZHOU Xin-Yu;YIN Zi-Yue;GAO Wei-Feng;TAN Gui-Sen;YI Yu-Gen(School of Computer and Information Engineering,Jiangzi Normal University,Nanchang 330022;Department of Applied Mathematics,Xidian University,Xi'an 710126;School of Software,Jiangri Normal University,Nanchang 330022)

机构地区：[1]江西师范大学计算机信息工程学院,南昌330022 [2]西安电子科技大学数学与统计学院,西安710126 [3]江西师范大学软件学院,南昌330022

出　　处：《计算机学报》2024年第7期1521-1546,共26页Chinese Journal of Computers

基　　金：国家自然科学基金(62366022,61966019,62276202,62062040);江西省自然科学基金杰出青年基金项目(20212ACB212003);江西省主要学科学术和技术带头人培养计划(20212BCJ23017);江西省自然科学基金项目(20232BAB202048)资助.

摘　　要：邻域拓扑是提高人工蜂群算法性能的一种有效手段.然而,现有相关工作主要是在种群层次上实现了单一邻域拓扑,这种方式忽略了不同类型的邻域拓扑能优势互补,使得算法性能还有一定局限性.为此,本文结合强化学习,提出在个体层次上实现多邻域拓扑。将种群中的个体视作智能体,设计了基于邻域拓扑的状态和动作,选用4种不同特征的邻域拓扑用于构建邻域候选池,之后采用Q-learning方法根据个体的奖励情况为其自适应选择不同的邻域拓扑.该方式相对于现有的单一邻域拓扑,更能充分发挥不同邻域信息对算法搜索的引导作用.在CEC2013和CEC2017两套测试集以及两个实际优化问题上进行了大量实验,与4种邻域相关ABC和4种知名改进ABC进行了性能对比,结果表明该算法的收敛精度和速度均有更好表现,可有效增强邻域人工蜂群算法的性能.Neighborhood topology is an effective way to enhance the performance of artificial bee colony(ABC)algorithm.However,for the existing related works,they mainly focus on a single neighborhood topology at the population level.This way overlooks the complementary advantages of different neighborhood topologies,and limits the algorithm performance to some extent,To remedy this issue,in this work,we propose the multiple neighborhood topologies mechanism at the individual level by combining the reinforcement learning.Specifically,the individuals in the population as considered as agents,and then the state and action of the agent are designed based on the neighborhood topology.After that,four different types of neighborhood topologies are used to construct the neighborhood candidate pool,and then the Q-learning method is employed to adaptively select different neighborhood topologies for the individuals according to their rewards.Compared with the existing methods of a single neighborhood topology,our approach can make full use of different neighborhood information in guiding the algorithm search.Extensive experiments are conducted on the CEC2013 and CEC2017 test suites,as well as two real-world optimization problems.Comparative results against four neighborhood-based ABC variants and other four well-known ABC variants demonstrate that our approach shows better performance in terms of both convergence accuracy and speed,and it can significantly improve the performance of neighborhood-based ABC.

关 键 词：群智能 人工蜂群 个体 强化学习 邻域拓扑 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]